Unbiasing One-vs-all When you train a one-vs-all multi-class classification, the rule of thumb is that for each class (e.g. class A) mark it as class 0 and others as class 1. Then you split the data as you wish and train the classifier. 
My question is, this creates a biased classifier: the number of samples of each class is not equal, the number of observations in class A is much smaller than all others. Should I take on measures to subsample all others or resample from class A to balance positive and negative observations for the classifier?
For example: take on a 5 class classification where for each class you have 5 observations. A one-vs-all means you mark for each class 5 positive examples and 20 negative examples.
 A: In such a case, you may have a biased estimator since data will be imbalanced as you mentioned. But this is ok as long as you are aware to the different mis-classification errors (and their costs) and their interpretations. All errors cost the same?
Yet, the question of whether to take care of the bias and how - depends on the interpretation of the final model.
If you want to remove the bias during training, this is not different than an imbalanced binary classifier. You can use proper overfitting and balancing strategies such as under/over-sampling, or setting sample/class weight to the estimator.
More important, one should also use the appropriate performance metrics to validate the model and investigate the misclassifications properly; balanced accuracy score, ROC AUC scores for each class ("one-vs-one"/"one-vs-all" and averaging by taking the support into account), plot AUCs for each class (TPR vs FPR, and precision-recall curves), and plot confusion matrix for each class. Although you're using "one-vs-all" strategy, validating the model as "one-vs-one" metrics will provide you additional validation on different biases.
